Refactor mkldnn-specific parts out of _inductor/overrides

[ngimel]

Currently a lot of mkldnn-specific functions and classes reside in _inductor/overrides.py under fairly generic names such as ConvUnary2d

pytorch/torch/_inductor/overrides.py

Line 84 in 3c637e8

class ConvUnary2d(nn.Conv2d):

and fuse_unary

pytorch/torch/_inductor/overrides.py

Line 699 in 3c637e8

def fuse_unary(gm: torch.fx.GraphModule):

. This is confusing, and code should be structured to make it clear when mkldnn specific passes happen. Moving these functions and classes to a separate file clearly marked as mkldnn would be helpful.
cc @ezyang @soumith @msaroufim @wconstab @bdhirsh @mlazos @voznesenskym @yanboliang @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @chunyuan-w @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @peterbell10 @desertfire, @EikanWang

[soumith]

yea this makes sense. @jgong5 and team should take a lead and quickly refactor this

[jgong5]

yea this makes sense. @jgong5 and team should take a lead and quickly refactor this

@soumith @ngimel Sure. Should be quick to do that.